Image attachments — hand a mockup to an agent

“Here’s the design, build it” needs the agent to actually see the design. Attachments save the image as a real file under the workspace’s attachments/ folder and hand the path to the CLI — Claude Code opens paths referenced in the prompt; Codex takes them as --image arguments.

Four ways to attach

1. 📎 on the agent input — pick image files (PNG, JPEG, GIF, WebP) from disk.
2. Paste into the input — copy any image, click the agent’s input, Ctrl+V. The chip appears with a thumbnail.
3. Send to agent — hover a pasted image on the canvas, or open the Screenshot pane, and click the ↗ send button. Pick an agent from the list; the image attaches there and the window comes to you. Screenshot sends include your annotations.
4. Drag it onto the window — grab any image on the canvas (pasted, or brought in with the left-toolbar Image tool) and drop it on an agent. While you drag, the window under your pointer lights up: a green 📎 Drop to attach ring means releasing attaches the image and brings that agent forward; a red Can’t view images ring means that agent’s CLI can’t see images — releasing there just moves the image, and a toast says why. Either way the image itself stays on the canvas.

Chips, pins, and sending

Attached images render as chips above the input — thumbnail, name, a pin, and a remove ✕. To remove one: click its ✕; or just press Backspace / Delete while that agent window is in front (removes the most recent attachment, as long as you aren’t mid-typing in the prompt); or click a specific chip (it gets an accent ring) and press Delete to remove that one. Send your message and the transcript notes 📎 N image(s) attached; unpinned chips clear with the send, pinned ones stay for the next iteration.

Good to know

• Attachment files persist under the workspace folder, so a pinned chip survives an app restart (the thumbnail falls back to a glyph until re-added).
• Keep images ≤5 MB each for Claude Code; that’s its documented per-image limit.